GraphRAG is an open-source data pipeline and research implementation from Microsoft for improving retrieval-augmented generation with structured knowledge extracted from unstructured text. It uses LLMs to transform source material into graph-oriented representations that can help models reason about entities, relationships, communities, and broader themes in private datasets.

Instead of relying only on conventional chunk retrieval, GraphRAG builds knowledge structures that can support questions requiring connections across many pieces of information. The repository includes indexing and query workflows, command-line tooling, configuration, prompt-tuning guidance, and migration notes. Because graph extraction uses LLM calls, the maintainers warn that indexing can be expensive and recommend starting with small datasets and tuning prompts for the target domain.

GraphRAG is aimed at researchers and developers working with complex corpora where ordinary semantic retrieval may miss global relationships or multi-hop context. The code is presented as a methodology and demonstration rather than a formally supported Microsoft product, making it especially useful for experimentation, evaluation, and building customized graph-based RAG systems.